🐛 Bug

When using weighed edges as input for GCNConv-Layer, I get some nans as results which do not make any sense. The problem is connected to edge_weights. When using edge_weights=None, the problem does not occur. Changing the values of x, edge_index and edge_weights leads to different values and different positions of nans. I provide the full setup to reproduce this error below:

To Reproduce

import torch.nn as nn
import torch.nn.functional as F
import torch_geometric
import torch_geometric.nn as gnn
torch.manual_seed(0)
gcn = gnn.GCNConv(in_channels=1, out_channels=4)

x = torch.randn(10, 1)
edge_index = torch.randint(0, 10, [2, 22])
edge_weight = torch.randn(22)
y = gcn(x=x, edge_index=edge_index, edge_weight=edge_weight)
`

Expected behavior

Environment

• OS: Linux
• Python version: Python 3.7.4
• Torch version: 1.3.0
• CUDA version: 9.0
• GCC version: 4.8.5

Additional context

The problem is based on issue 707